This proposal requests funding to continue research aimed at developing statistical models and methods to help medical oncologists find the best way to treat cancer patients with pharmaceutical agents. Understanding the many sources of between-patient heterogeneity is key to individualizing therapy. Individualizing therapy for cancer patients is particularly important, because of the relatively narrow therapeutic window of most cancer chemotherapy. Our proposed statistical models will aid the discovery of sources of variation in a patient's response to therapy, allowing individualizing anticancer chemotherapy. The overall hypothesis of this application is that understanding the pharmacokinetics (PK), pharmacodynamics (PD), and pharmacogenetics (PGx) of anticancer agents is important for evaluating efficacy and determining how best to use such agents clinically. New statistical methodology will help synthesize this information more efficiently to improve treatment outcomes for cancer patients. Our proposal has four specific aims. (1) Develop a decision-theoretic approach to facilitate dose individualization. Our approach combines patient-specific PK information with PK data from other patients treated at a range of doses to help determine an optimal dose for the current patient. (2) Build coherent statistical models to learn about the dependence of PK summaries and related SNPs. The methods will foster discovery through joint models for PK of a drug, its metabolites, and related genotype information. (3) Develop Bayesian nonparametric models for repeated data nested within repeating cycles. These data often arise in clinical studies, where interest may concern studying the within-patient correlation structure across cycles or learning about patient-specific characteristics that influence outcomes while accounting for the repeated-repeated measurement structure. (4) Construct a framework to model PK and PD data simultaneously, considering PK and PD responses as functional responses, rather than focusing on a few low-dimensional summaries. We will develop a joint probability model that will implement regression for function-valued data, considering PK as the explanatory variable and the longitudinal PD response as the outcome. Underlying methodological themes are the efficient use of all data that can be collected in the course of a study, joint inference on unknown quantities with models that appropriately propagate uncertainties, and an attempt to cast the scientific questions of interest as statistical inference questions and, where applicable, as decision problems.